Résumé: This study presents a novel algorithm for classifying pulmonary diseases using lung sound signals by integrating Variational Mode Decomposition (VMD) and the Constant-Q Transform (CQT) within a pre-trained AlexNet convolutional neural network. Breathing sounds from the ICBHI and KAUHS databases are analyzed, where three key intrinsic mode functions (IMFs) are extracted using VMD and subsequently converted into CQT-based time-frequency representations. These images are then processed by the AlexNet model, achieving an impressive classification accuracy of 93.30%. This approach not only demonstrates the innovative synergy of CQT-VMD for lung sound analysis but also underscores its potential to enhance computerized decision support systems (CDSS) for pulmonary disease diagnosis. The results, showing high accuracy, a sensitivity of 91.21%, and a specificity of 94.9%, highlight the robustness and effectiveness of the proposed method, paving the way for its clinical adoption and the development of lightweight deep-learning algorithms for portable diagnostic tools.

Résumé: In lung sound classification using deep learning, many studies have considered the use of short-time Fourier transform (STFT) as the most commonly used 2D representation of the input data. Consequently, STFT has been widely used as an analytical tool, but other versions of the representation have also been developed. This study aims to evaluate and compare the performance of the spectrogram, scalogram, melspectrogram and gammatonegram representations, and provide comparative information to users regarding the suitability of these time-frequency (TF) techniques in lung sound classification. Lung sound signals used in this study were obtained from the ICBHI 2017 respiratory sound database. These lung sound recordings were converted into images of spectrogram, scalogram, melspectrogram and gammatonegram TF representations respectively. The four types of images were fed separately into the VGG16, ResNet-50 and AlexNet deep-learning architectures. Network performances were analyzed and compared based on accuracy, precision, recall and F1-score. The results of the analysis on the performance of the four representations using these three commonly used CNN deep-learning networks indicate that the generated gammatonegram and scalogram TF images coupled with ResNet-50 achieved maximum classification accuracies.

Résumé: The acquisition of Breath sounds (BS) signals from a human respiratory system with an electronic stethoscope, provide and offer prominent information which helps the doctors to diagnosis and classification of pulmonary diseases. Unfortunately, this BS signals with other biological signals have a non-stationary nature according to the variation of the lung volume, and this nature makes it difficult to analyze and classify between several diseases. In this study, we were focused on comparing the ability of the extreme learning machine (ELM) and k-nearest neighbour (K-nn) machine learning algorithms in the classification of adventitious and normal breath sounds. To do so, the empirical mode decomposition (EMD) was used in this work to analyze BS, this method is rarely used in the breath sounds analysis. After the EMD decomposition of the signals into Intrinsic Mode Functions (IMFs), the Hjorth descriptors (Activity) and Permutation Entropy (PE) features were extracted from each IMFs and combined for classification stage. The study has found that the combination of features (activity and PE) yielded an accuracy of 90.71%, 95% using ELM and K-nn respectively in binary classification (normal and abnormal breath sounds), and 83.57%, 86.42% in multiclass classification (five classes).

Résumé: In this paper, we propose a novel approach for Phonocardiogram (PCG) signal classification using a BiLSTM model with Mel Frequency Cepstral Coefficients (MFCC) features extracted from short PCG segments. The issue of handling PCG signals of varying lengths is addressed by segmenting the audio signal, allowing for feature extraction and organization with a fixed dimension that is compatible with the input layer of the BiLSTM model. Our approach achieves state-of-the-art performance while utilizing only two BiLSTM layers, making it an efficient and lightweight model for embedded applications. The combination of MFCC features, a handcrafted feature extraction method, with a BiLSTM architecture, addresses the issue of feature engineering for improving PCG signal classification performance. Our study is the first work in the literature to explore the potential benefits of using MFCC features with a BiLSTM model for PCG signal classification. The proposed approach has the potential to significantly impact the healthcare industry by improving the accuracy and efficiency of PCG signal classification, aiding in earlier diagnosis and treatment.

Résumé: Breathing sounds are a rich source of information that can assist doctors in diagnosing pulmonary diseases in a non-invasive manner. Several algorithms can be developed based on these sounds to create an automatic classification system for lung diseases. To implement these systems, researchers traditionally follow two main steps: feature extraction and pattern classification. In recent years, deep neural networks have gained attention in the field of breathing sound classification as they have proven effective for training large datasets. In this study, we conducted a comparison of two versions of the VGG16-based deep learning model for breathing sound classification using Gammatonegrams as input. We implemented two extensions of the VGG16 model - one executed from scratch and the other based on a pretrained VGG16 model using transfer learning. We processed digital recordings of cycle-based breathing sounds to obtain Gammatonegrams images, which were then fed as input to the VGG16 network. In addition, we performed data augmentation in our experiments using audio cycles from the ICBHI database to evaluate the performance of the proposed method. The classification results were obtained using the Google Collaboratory platform.

Résumé: Many classification algorithms have been implemented to differentiate between different pulmonary diseases. Recently, machine learning techniques have used for lung sound classification, and have particularly focused on deep neural networks, which appear advantageous with large training datasets. In this paper, intending to provide a fully automatic classification system, we propose an alternative representation of input data called Gammatonegrams. Our approach was implemented on two different deep neural network architectures - VGG16 and ResNets for pulmonary pathologies classification. The ICBHI database was chosen as input for pulmonary conditions classification into- healthy, chronic and non-chronic. The results show that the two architectures gave an accuracy of 67.97 % and 60.80% for VGG16 and ResNet-50 respectively. Our results provide initial evidence that in the gammatonegram based classification of pulmonary conditions, the deep neural networks, can achieve significant accuracy.

Résumé: During the diagnosis of pulmonary diseases, the doctor listens to the lung sounds on the patient’s chest using a traditional stethoscope for the earlier detection of abnormal respiratory sounds. Currently, due to developments in the digital technology domain, an electronic stethoscope can instead be used to acquire respiratory sounds and save these as data for further processing and analysing. Many algorithms for automatic classification have been implemented to distinguish between several lung diseases. In this study, we implemented a deep residual network (ResNet) model with three different architectures based on different numbers of layers (ResNet-50/101/152) for the classification of pulmonary pathologies. To evaluate this method, we used it for the analysis of gammatonegrams, which transform lung sounds from onedimensional to two-dimensional representations. The image outputs obtained with the gammatonegram are fed as inputs to the Three-ResNet architecture. The ICBHI database was used to classify three types of pulmonary conditions, namely, healthy, chronic obstructive pulmonary disease (COPD) and pneumonia conditions. The results showed that, ResNet-50, ResNet-101 and ResNet-152 presented accuracies of 90.37%, 89.79% and 67.57%, respectively. Therefore, our results demonstrate that the residual networks can achieve significant accuracy for the classification of these three types of pulmonary conditions.

Résumé: The traditional methods used by researchers when implementing breathing sounds classification systems involve two main steps - feature extraction and pattern classification. In recent years, the topic of interest in the field of breathing sound classification focuses on the use of deep neural networks, which have been proven to be effective for training large datasets. In this paper, we conducted a comparative study of three deep neural network architectures, the VGG16, ResNet-50, and GoogLeNet for breathing sounds classification. Digital recordings of cycle-based breathing sounds from the ICBHI database are processed to obtain gammatonegram images that are fed as an input to these three networks. The classification results, executed on the Google Colaboratory platform, indicated that these three networks yielded accuracies of 62.50%, 62.29%, and 63.69% respectively. Hence, the results provide initial evidence that GoogLeNet can significantly improve the accuracy and outperformed the VGG16 and Resnet-50 in our application.

Résumé: Breathing sounds contain prominent information that can aid doctors to diagnose pulmonary pathologies in a non-invasive way. Based on these sounds, we can establish many algorithms to develop an automatic classification system that could be used to categorize lung diseases. The traditional methods used by researchers when implementing these systems involve two main steps – feature extraction and pattern classification. In recent years, the topic of interest in the field of breathing sound classification focuses on the use of deep neural networks, which have been proven to be effective for training large datasets. In this paper, we conducted a comparative study that involves the use of Gammatonegrams as input to the two versions of VGG16-based for breathing sounds classification. We implemented two extensions of the Visual Geometry Group 16 (VGG16) deep learning model, the first VGG16 was executed from scratch, and the second was based on a VGG16 pre-trained model (transfer learning). Digital recordings of cycle-based breathing sounds are processed to obtain Gammatonegrams images which are fed as input to the VGG16 network. Using audio cycles from the ICBHI database for both extensions, we additionally performed data augmentation in our experiments to measure the performance of the proposed method. The classification results, are executed on the Google Colaboratory platform.

Résumé: In this paper, a statistical analysis is performed for many parameters such as Harmonic to Nois Ratio (HNR),Pitch and Amplitude Perturbation (jitter and shimmer) was applied to compare between them in different cases from the respiratory signal. COPD and Pneumonia are a kind of pulmonary diseases which affects the respiratory signal. Respiratory sounds is non-stationary, non-linear and complex signals. The purpose of this work is acoustic features extraction using PRAAT software v6.0.43 from the respiratory signal. In order to evaluate these features, the mean and standard deviation are calculated and presented. However, in our study we found the Intensity value to be 66.42, 76.58 and 71.96 for the healthy, COPD and Pneumonia patients respectively, The Jitter(s) values obtained for the healthy, COPD and Pneumonia patients as 101.03, 174.40 and 173.45 respectively. Our results show the significantly increased values in COPD and Pneumonia patients compared to healthy subjects.

Résumé: The acquisition of Breath sounds (BS) signals from a human respiratory a system with an electronic stethoscope, provide and offer prominent information which helps doctors to diagnose and classification of pulmonary diseases, hence in this paper we propose a prototype to classify lung sounds using a raspberry pi and an electronic stethoscope such as Littman Or PCP.

Résumé: Over the years, breath sounds signals have been used to diagnose pulmonary pathologies. This signal contains valuable important information, it helps doctors to detect and classify several diseases such as (Asthma, Pneumonia, and COPD….). Unfortunately, the nature of these signals is non-linear and non-stationary, which makes it difficult to analyze and identify different types of sounds. To address this problem, there is a signal processing algorithm in the time-frequency domain that has the same nature as these signals namely EMD (Empirical Mode Decomposition), which was used in this work to analyse the breath sounds signal. The decomposition of the signals by the EMD method produces many Intrinsic Mode Functions (IMF), from each IMF we extracted two features (Energy, Entropy) and combined them for the first classification stage, and separate them for the second classification stage. The study has found that the combination of features (Energy and Entropy) yielded an accuracy of 84.61% using K-nn classifier, and the maximum classification accuracies obtained of the energy and entropy features with the same classifier were 98.33 % and 76.66%, respectively.